There’s a story about Claude that’s been everywhere this week: it identified a writer from 125 unpublished words. It’s been read as a privacy story. For people who write, the more useful read is that it’s a craft story.
Earlier this week, journalist Kelsey Piper, writing for The Argument, pasted 125 words of an unpublished political column into Claude Opus 4.7 and got her own name back. She’d logged out, tested via the API, run it again on a friend’s laptop. She varied the genre too: a school progress report about her child’s Pokémon essays, an unpublished review of a 1942 wartime comedy. Claude named her every time. ChatGPT and Gemini failed the same task. (Read her full account.)
Privacy is the frame the news has settled on. There’s a different frame to read it through if you write.
What the Piper experiment rules out
The Piper experiment matters because of how she designed it. The four methods she used systematically rule out every alternative explanation except stylometry.
Each method closed off a different escape route for the model. Logging out and switching to incognito killed the obvious “Claude knows me from my account” answer. Switching to the raw API ruled out browser fingerprinting. Repeating the test on a friend’s laptop ruled out IP-based identification. By the time those four were exhausted, the only remaining channel through which the model could know her was the prose itself.
Then she varied the genre. A political column might overlap with her public corpus. A school progress report about a child’s Pokémon essays does not. A review of a 1942 wartime comedy is not in her published register at all. The model still returned her name. That detail is the one that does the work. It means Claude isn’t matching topics or subject matter. It’s reading the way her sentences are built.
Which is stylometry. And stylometry is now a working part of frontier language models, whether anyone intended that or not. ChatGPT and Gemini, run on the same task, guessed wrong. The capability is uneven across the field. The fact that one model already has it changes the shape of the question for writers anyway.
What does the privacy framing of this story miss?
The privacy reading treats the writer as the subject of surveillance. The craft reading treats prose itself as a measurable artefact, which is the more important implication if you write for a living.
The privacy reading of the Piper experiment is real and worth taking seriously. Models can identify writers from short, unpublished, off-genre prose, and the threshold will only drop as models improve. Fair concern. Useful coverage.
But the privacy reading treats the writer as the subject of surveillance. For working writers, that’s the secondary problem. The primary problem is what the experiment proves about prose itself.
Voice is real. Not in the way we wave at it when we talk about craft over a drink. In the boring measurable way. You can be identified from a few hundred words of your own writing. So can I. So can the writer you most admire.
Which I think most working writers already suspected. But it’s one thing to suspect and another to see Claude do it.
If voice is measurable, voice is preservable. And if voice is preservable, the question stops being aesthetic taste and becomes craft engineering. That’s the story that matters if you’re drafting fiction or long-form non-fiction with AI. Privacy is a perimeter problem. Voice is a craft problem. The craft problem is the one nobody else in this discourse is writing about.
Why is voice a fingerprint, not an aesthetic?
Voice is a multi-axis stylometric fingerprint that includes word choice, sentence shape, paragraph structure, rhythm, and dozens of other axes. It is distinct enough for a frontier language model to identify a writer from 125 words of off-genre prose.
Most AI writing tools treat voice as an aesthetic variable. You pick formal or casual, literary or commercial, sparse or maximalist. The dropdown is the voice. This framing has been slowly wrong for a while, but the Piper experiment makes it provably wrong.
Voice is a multi-axis fingerprint. It includes word choice, sentence shape, the ratio of declaratives to dependent clauses, paragraph openings, paragraph closings, the rhythm of how sentences shorten or lengthen under emotional pressure, the writer’s relationship to abstraction versus concrete imagery, the placement and frequency of parenthetical asides. Claude Opus 4.7 can read enough of those axes from 125 words to narrow a candidate down to a single writer in a public corpus. That’s stylometry, and stylometry is now a working part of frontier language models, whether anyone intended that or not.
Treat that finding as the design constraint it is. If voice is measurable, it’s preservable. And once it’s a measurable constraint, the tools that handle it well will separate hard from the tools that pretend voice is a setting.
This is the same fingerprint we wrote about in how to find your writing voice. The Piper experiment is the proof that the fingerprint is real and measurable, not just rhetorically convenient.
What happens when AI writes “in your voice”?
AI writing tools that don’t constrain voice produce technically competent prose that no longer matches the writer’s fingerprint. The drift is measurable, demonstrable, and now provable through the same stylometric analysis the Piper experiment exposed.
I tried Piper’s experiment on a passage from my own novel a couple of nights ago. I asked Claude to describe the voice in 200 words I’d written. The description was startlingly specific. It told me my prose has an instinct for off-key similes that collapse mid-air, and quoted one back at me: a character’s hand gesture indicating “either a volcano or aggressive udder milking.” I had not noticed that about my own writing.
Then I asked it to generate a new scene in that voice. The result was technically competent and not mine. The shortening was gone. Adverbs were back. A kind of generic literary cadence had moved into the middle of the prose like a stranger.
Which is, I suspect, what most writers mean when they say AI drafts don’t sound like them. The feeling isn’t imaginary. The fingerprint is gone, and now we have a way to see it going.
How should writers think about this going forward?
Voice is a craft engineering problem now, not a vibes argument. The category of AI writing tools worth using is the one that treats voice as a binding constraint on every output, not as a setting to be specified once and forgotten.
Two practical implications.
First, voice is no longer something writers have to argue for. It’s measurable. When a tool sands the voice out of your draft, you can demonstrate it, and you should expect the tools you use to take that demonstration seriously.
Second, the question of “how do I keep AI from flattening my voice” stops being a vibes argument and becomes a craft problem. Solvable. Worth the effort. Different tools approach it differently, with different rates of success. The category that matters going forward is whether a tool treats voice as a constraint or as a setting. Constraints govern every output. Settings get ignored after the third paragraph.
Privacy is a real concern in the Piper story. Voice as a measurable craft constraint is the bigger one for anyone making things from prose. The next year of AI writing tools will be sorted along that axis, whether the tools know it or not. Writers will sort them.